Data Center Control Against Sub-Synchronous Resonance: A Data-Driven Approach

TL;DR

This paper introduces a data-driven impedance modeling and control method to assess and mitigate sub-synchronous resonance risks in grid-connected data centers, enhancing grid stability and safety.

Contribution

It develops a workload-dependent impedance modeling approach and a preventive control mechanism for early SSR risk detection and workload management.

Findings

Impedance characteristics vary significantly with workload.

The proposed method effectively predicts resonance risks.

Early warnings improve safety margins with minimal workload adjustments.

Abstract

Data centers host a variety of essential services such as cloud computing and artificial intelligence. Electric grid operators, however, have limited knowledge of the reliability risks of data center interconnection due to their unique operational characteristics. An emerging concern is the sub-synchronous resonance (SSR) which refer to unexpected voltage/current oscillations at typical frequencies below 60/50 Hz. It remains unknown whether and how the interactions between data centers and the grid may trigger resonances, equipment damages, and even cascading failures. In this paper, we focus on grid-connected data centers that draw electricity from the grid through power factor correction (PFC) converters. We conduct two-tone frequency sweep to investigate the data centers' impedance characteristics, i.e. magnitude and phase angle variations over frequencies, and showcase their deep dependence on compute workloads. The impedance modeling provides a direct approach to evaluating SSR risks and enable a cooperative mechanism to alarm and avoid resonance-prone situations. Building upon the impedance, a data-driven preventive controller is then established to raise early warnings of risky operation and suggest flexible workload management according to the given grid conditions. Through case study, we demonstrate how to use impedance to understand the unexpected interactions. Data-driven impedance is validated to show decent performance in capturing the unique impedance dips and tracking the impedance variations across a range of workload scenarios. The early warning and preventive control approaches are further effective to improve the safety margins with minimal workload rescheduling. The key findings of this work will provide valuable insights for grid operators and data center managers, and support preparation for future scenarios involving large-scale data center integration.